Building with AI agents. Notes on Claude Code, Codex, product engineering, and what it means to ship faster
Joined February 2021
- Tweets666
- Following313
- Followers21K
- Likes680
Pinned Tweet
1/
For three years I wrote about crypto under @cryptoPothu.
NFTs, DeFi, on-chain analysis, whale wallets, where the money was moving.
I got a lot right. I got some things badly wrong.
Both taught me the same thing: how to find edges in noisy markets.
I’m taking that somewhere more interesting.
Pothu retweeted
90% of the "AI startups" in your feed will be a feature inside Claude or ChatGPT within 18 months.
The 10% that survive own one of three things. Everything else is a system prompt waiting to be absorbed.
The uncomfortable part: most of these products are genuinely good. Clean UI, real users, decent revenue.
That's not enough anymore. "we do one thing well" is the exact description of a feature, not a company.
Survivor trait #1: proprietary data the model can't get anywhere else.
Not "we use AI on data." Data that only exists because of something you built. The labs can copy your interface in a weekend. They can't copy a dataset they don't have.
Survivor trait #2: a regulated or high-trust workflow.
Healthcare, legal, finance, anything where being wrong has consequences and someone has to be accountable. A chatbot can't sign off. That liability moat is real and the labs don't want it.
Survivor trait #3: a brand and a distribution channel that existed before the AI feature.
The thing nobody building wants to hear: distribution was always the hard part. Agents made building cheap and did absolutely nothing for getting people to care.
I build with agents every day. I can spin up a working v1 of almost any "AI startup" I see posted, in a weekend.
That's not a flex. It's the whole problem. If I can clone your product in two days, your product was never the moat.
"Saas is dead" was always lazy. Thin-wrapper saas is dead. The rest got harder, not deader.
The build stopped being the bottleneck. Distribution, trust, and proprietary data are the entire game now. They always quietly were.
90% of the "AI startups" in your feed will be a feature inside Claude or ChatGPT within 18 months.
The 10% that survive own one of three things. Everything else is a system prompt waiting to be absorbed.
The uncomfortable part: most of these products are genuinely good. Clean UI, real users, decent revenue.
That's not enough anymore. "we do one thing well" is the exact description of a feature, not a company.
Survivor trait #1: proprietary data the model can't get anywhere else.
Not "we use AI on data." Data that only exists because of something you built. The labs can copy your interface in a weekend. They can't copy a dataset they don't have.
Survivor trait #2: a regulated or high-trust workflow.
Healthcare, legal, finance, anything where being wrong has consequences and someone has to be accountable. A chatbot can't sign off. That liability moat is real and the labs don't want it.
Survivor trait #3: a brand and a distribution channel that existed before the AI feature.
The thing nobody building wants to hear: distribution was always the hard part. Agents made building cheap and did absolutely nothing for getting people to care.
I build with agents every day. I can spin up a working v1 of almost any "AI startup" I see posted, in a weekend.
That's not a flex. It's the whole problem. If I can clone your product in two days, your product was never the moat.
"Saas is dead" was always lazy. Thin-wrapper saas is dead. The rest got harder, not deader.
The build stopped being the bottleneck. Distribution, trust, and proprietary data are the entire game now. They always quietly were.
Pothu retweeted
For 6 months my stack was Gemini for research, Claude for UI, Codex for dev. After one day with Fable 5, I think that's over.
Here's the three-test experiment that convinced me:
1. One vague sentence.
My CI was red this morning, and the PR that broke it was written by Opus 4.8 in the same chair. I typed "lint was failing in recent PR" and nothing else. Fable dug up the failed run, read the log, fixed both files, opened the PR. 3 minutes. Didn't ask me a single thing. With Codex I'd have pasted the run link and file paths first.
2. Real work.
First PR of my project's new phase: an architecture decision doc, data-source research, typed contracts in Python and Zod, a market classifier, 22 tests. One shot, all gates green, 236 backend tests passing. It even rejected a data source on ToS risk I hadn't asked about.
3. Felt rude.
I spawned 4 parallel Fable reviewers and pointed them at the diff Fable had just written. They found 8 issues. Best catch was a copy-paste prefix list that had already drifted, with the exact failure mode predicted. Fable wrote that bug 30 minutes before another Fable explained why it would happen.
The warts, because nobody posts these: the safety classifier gating shell commands went down twice on me ("claude-fable-5 is temporarily unavailable"), and some topics reroute to Opus 4.8 under the hood (under 5% of sessions per Anthropic, with a banner). You're occasionally paying Fable rates for Opus answers. And the rates are real: $10 in, $50 out per million tokens, about double GPT-5.5.
Day-one scorecard: 3 PRs opened, 1 merged, 22 new tests, one self-caught bug, two infra outages. Next week I hand it an entire project phase. If the honeymoon dies the way Devin's did, I'll post that too.
For 6 months my stack was Gemini for research, Claude for UI, Codex for dev. After one day with Fable 5, I think that's over.
Here's the three-test experiment that convinced me:
1. One vague sentence.
My CI was red this morning, and the PR that broke it was written by Opus 4.8 in the same chair. I typed "lint was failing in recent PR" and nothing else. Fable dug up the failed run, read the log, fixed both files, opened the PR. 3 minutes. Didn't ask me a single thing. With Codex I'd have pasted the run link and file paths first.
2. Real work.
First PR of my project's new phase: an architecture decision doc, data-source research, typed contracts in Python and Zod, a market classifier, 22 tests. One shot, all gates green, 236 backend tests passing. It even rejected a data source on ToS risk I hadn't asked about.
3. Felt rude.
I spawned 4 parallel Fable reviewers and pointed them at the diff Fable had just written. They found 8 issues. Best catch was a copy-paste prefix list that had already drifted, with the exact failure mode predicted. Fable wrote that bug 30 minutes before another Fable explained why it would happen.
The warts, because nobody posts these: the safety classifier gating shell commands went down twice on me ("claude-fable-5 is temporarily unavailable"), and some topics reroute to Opus 4.8 under the hood (under 5% of sessions per Anthropic, with a banner). You're occasionally paying Fable rates for Opus answers. And the rates are real: $10 in, $50 out per million tokens, about double GPT-5.5.
Day-one scorecard: 3 PRs opened, 1 merged, 22 new tests, one self-caught bug, two infra outages. Next week I hand it an entire project phase. If the honeymoon dies the way Devin's did, I'll post that too.
6/
There IS one real 4.8 regression nobody's posting about: its prompt-injection robustness got worse than 4.7.
That's a number in a doc. "feels dumber after lunch" is a vibe. Notice which one the timeline picked to be mad about.
You can ask one AI company CEO one question and they have to answer honestly.
Who are you picking?
Sam Altman
Dario Amodei
Elon Musk
Sundar Pichai
And what’s your question?
7/ Last one: how a model fails matters more than how often.
4.8's best upgrade for me isn't a score, it's that it stops and says "this part is shaky" instead of confidently rewriting a working file. An eval counts the pass. It can't count the 40 minutes a confident wrong answer costs you.